Regensburg 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 36: Mechanical Properties

MM 36.7: Vortrag

Freitag, 21. März 2025, 12:00–12:15, H23

Transferable machine learning interatomic potential for Au nanoparticles — •Jovana Vlahović, Cem Sevik, and Milorad V. Milošević — University of Antwerp, Groenenborgerlaan 171, Antwerp, Belgium

Effective molecular dynamics and Monte Carlo simulations fundamentally depend on the accuracy of interatomic potentials, which define the potential energy surface as a function of atomic positions. This precision is essential for metallic nanoparticles (NPs), whose small size and diverse morphologies demand robust modelling. While classical empirical potentials offer computational efficiency, they often lack the necessary accuracy and density functional theory (DFT) calculations, though highly accurate, are computationally prohibitive for NPs with thousands of atoms. This work introduces a machine learning-based interatomic potential for Au NPs, trained on a compact dataset of bulk structures, surface slabs and NPs containing up to 55 atoms. Our training dataset includes a diverse set of atomic configurations with corresponding properties obtained from ab initio molecular dynamics calculations, which we use to train a Gaussian Approximation Potential (GAP) via Gaussian Process Regression. Parameter optimization was performed to maximize model accuracy, and LAMMPS validation tests demonstrated GAP performance against DFT benchmarks. In addition, transferability tests on larger NPs of various shapes reveal our GAP's robustness beyond the training set. Comparisons with an existing GAP model and the universal MACE potential underscore our model's improved accuracy and generalizability for Au NPs.

Keywords: Gold nanoparticles; Machine learning interatomic potentials; Gaussian Approximation Potential; Molecular dynamics